SHIPPED
AI institutionalization
Bringing AI into a traditional insurance brokerage — AI-assisted requirements analysis and workflow optimization, cutting work-hours by ~60%. Tested first in my own lab (Hermes), then shipped company-wide.
Context
IT project manager at a Taiwan insurance brokerage, responsible for coordinating vendors and modernizing internal workflows.
Problem
A traditional brokerage runs on high-frequency manual work. "Use AI" is easy to say and hard to make stick — output has to be trustworthy and on-brand, not a novelty.
Constraints
Non-technical users. Brand consistency required. Output quality had to be reliable enough to trust in daily operations.
Approach
Multi-model routing (a strong model for architecture decisions, a fast one for code generation, cheap ones for batch tasks) to balance quality and cost. The design system was documented and injected into the AI tooling so generated front-ends matched brand rules automatically. A core discipline carried over from my lab: no agent verifies its own work — verification is always cross-agent. This was proven first in a personal testbed (Hermes) before being rolled out company-wide.
Outcome
~60% reduction in work-hours on targeted workflows. AI-assisted requirements analysis became a standard part of the process.
Stack
Claude API · multi-model routing · design-system injection · RAG